Research map / Vision-language and multimodal / CLIP and zero-shot
PACTran: PAC-Bayesian Metrics for Estimating the Transferability of Pretrained Models to Classification Tasks
Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, Radu Soricut
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PT-Gauss grid instead of LEEP | transferability metric · NeuCRaB, N/K=2 avg | Kendall-Tau correlation: 0.083 -> 0.680 | Improved |
| use GMM assignments (N-LEEP) instead of LEEP | LEEP · NeuCRaB, N/K=2 avg | Kendall-Tau correlation: 0.083 -> 0.401 | Improved |
| use PT-Gauss fix instead of LINEAR-VALID | transferability metric · NeuCRaB, N/K=10 avg | Kendall-Tau correlation: 0.480 -> 0.641 | Improved |
| add flatness regularizer to linear classifier loss (PT-Gauss fix vs LINEAR) (ablation) | LINEAR · NeuCRaB, N/K=2 avg | Kendall-Tau correlation: 0.139 -> 0.654 | Improved |
| use PT-Gauss fix instead of PT-Gauss grid when K is small | PT-Gauss · NeuCRaB, N/K=5 avg | Kendall-Tau correlation: 0.641 -> 0.673 | Improved |
| use PT-Gauss grid instead of fix at N/K=2 | PT-Gauss · NeuCRaB, N/K=2 avg | Kendall-Tau correlation: 0.654 -> 0.680 | Improved |
| use PACTran-Dirichlet/Gamma with GMM assignments instead of N-LEEP | N-LEEP · NeuCRaB, N/K=10 avg | Kendall-Tau correlation: 0.497 -> 0.477 | No clear effect |
| use LINEAR-VALID instead of LINEAR for small K | LINEAR · NeuCRaB, 2-9 classes, N/K=10 | Kendall-Tau correlation: 0.109 -> 0.123 | Mixed |
| use PT-Gam instead of LEEP | LEEP · OKVQA, N=200 Total | Kendall-Tau correlation: 0.508 -> 0.532 | Improved |
| use PT-Gauss grid instead of LINEAR-VALID | LINEAR-VALID · OKVQA, N=200 Total | Kendall-Tau correlation: 0.528 -> 0.582 | Improved |
| use PT-Gauss fix instead of penultimate feature extraction cost | feature extraction · NeuCRaB, N/K=10, 100+ classes | GFLOPS: 3.88E3 -> 6.45E1 | Improved |
| balance flatness and empirical-risk terms (ratio about 1.0) (ablation) | PT-Gauss · NeuCRaB VTAB | Kendall-Tau correlation | Improved |
Closest papers by meaning
- LogME: Practical Assessment of Pre-trained Models for Transfer Learning (ICML 2021 · similarity 0.86)
- A Bayesian Model Selection Criterion for Selecting Pretraining Checkpoints (ICML 2025 · similarity 0.86)
- Implicit Modeling for Transferability Estimation of Vision Foundation Models (NeurIPS 2025 · similarity 0.86)
- Fast and Accurate Transferability Measurement by Evaluating Intra-class Feature Variance (ICCV 2023 · similarity 0.86)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.86)
- Understanding prompt engineering may not require rethinking generalization (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
